tropism v2026.11.2: an MCP server that answers "does this change break a rule?"

Contents

The pre-commit hook tells you a change broke an architecture rule after you’ve written it. The new MCP server lets a coding agent ask before it writes the import.

tropism is a dependency analyzer for polyglot repositories. It finds import cycles, manifest problems and duplicated packages, and it enforces the architecture rules you’ve written in a tropism.toml at the repository root, across Go, JavaScript/TypeScript, Rust, C#, Python, Ruby, Java, Swift and C++. It never runs a package manager or the code it analyzes, so it works on a fresh checkout with no toolchain and no network. If you’ve wanted ArchUnit-style rules in a repository that isn’t just Java, that’s the niche. I wrote about why it’s built that way in Architecture the Agentic Loop Can Actually Check.

v2026.11.2 was tagged on 2 November. It has 8 commits and touches 22 files (1,464 insertions, 89 deletions), and almost all of that is one feature.

The headline is tropism-mcp, an MCP server that gives an agent three tools, over stdio:

  • tropism_check evaluates the rules against a list of files. It’s the same check the pre-commit hook runs, but asked before the commit. It reports only violations the listed files introduce, and returns a gate of pass, fail or unavailable (no ruleset to evaluate). fail means the hook would block the commit. Each finding carries the reason you wrote for the rule.
  • tropism_rules lists the ruleset with a verdict on each rule over the whole repository: satisfied, violated (with a count and examples), or stale, meaning the rule matched nothing and so protects nothing.
  • tropism_summary returns whole-repository counts per check and per project, plus which checks couldn’t run and why. It returns no individual findings, so it’s cheap enough to call first.

The part I care about most is that none of these can say “clean” when they mean “didn’t run”. Every response states which checks didn’t run, because an agent can’t infer that a zero means “never ran”. Lists are capped (50 findings per tropism_check call by default, with limit and offset for paging) and say when they were truncated.

The server is a thin adapter over the same pipeline the CLI uses, so there’s one implementation of each check, not two. That was the point of the design in design/05-interfaces.md: whatever the CLI can answer, the server answers the same way.

A caveat on how much to read into this. The server is a day old, and I’ve checked it against its own stdio test and the code, not against weeks of real agent sessions. Whether agents ask tropism_check at the right moment is the thing I still have to find out.

The README says only to put tropism-mcp on your PATH, so here’s what the code implies. The server takes no arguments, and each tool analyzes the server’s working directory unless the call passes a path. Start it from the repository root, which is what an MCP client configured per project normally does. In Claude Code that’s claude mcp add tropism -- tropism-mcp, run in the repository.

The README now has a Windows install that needs no admin rights. It downloads the latest release’s zip into %LOCALAPPDATA%\Programs\tropism, prints the zip’s SHA256 so you can compare it with the line in SHA256SUMS, expands it, and adds the directory to your user PATH. The binaries are unsigned, so SmartScreen or an application-control policy may still object.

Three changes are about how releases get built. Two of them might touch you.

Apple Silicon releases are now built on macos-15, which comes from syncing with release-kit, the shared release workflow my Rust tools use. The CalVer micro number is now one past the highest tag this month rather than a count of this month’s tags. The count had already gone wrong: November’s first tag was v2026.11.1, so counting gave 1 and make release named the existing v2026.11.1 again. And the internal crate dependencies now use >= requirements, so any tag resolves, which matters if you build from a tagged checkout.

The last commit is a record rather than a change: the design notes now list what moving from dist to release-kit dropped. The shell and PowerShell installers, the npm package and build-provenance attestation are gone. If you were installing through any of those, use one of the routes below.

From the README, on macOS or Linux with Homebrew:

brew trust --formula grahambrooks/tropism/tropism
brew tap grahambrooks/tropism https://github.com/grahambrooks/tropism
brew install grahambrooks/tropism/tropism

After that first tap, brew upgrade picks up new releases like any other formula. The README explains the first two lines: the formula lives in the repository rather than a separate tap, and Homebrew 6 refuses to tap an untrusted third-party tap, so the trust has to be granted first, with --formula. Older Homebrew has no brew trust and needs only the last two lines.

On any platform with bx, a tool of mine that downloads a GitHub project’s release archive for your machine and runs it:

bx grahambrooks/tropism check

Or take the archive for your platform from the releases page (Linux gnu x86-64 and arm64, Linux musl x86-64, macOS Intel and Apple silicon, Windows) and put tropism and tropism-mcp on your PATH. Every release carries a SHA256SUMS. On Windows, re-running the PowerShell lines from the README upgrades in place.

To build from source:

cargo build --release
./target/release/tropism analyze /path/to/repo

tropism isn’t on crates.io yet. Publishing is wired up but deliberately manual, because a version there can be yanked but never reused.

If you only use the CLI and the hook, this release changes little for you beyond the Windows instructions. If you work with coding agents, tropism-mcp is the reason to upgrade.